Auto mode becomes Claude Code's default, and a CVSS 9.4 GitHub Actions CVE

TLDR;

Auto mode becomes the default in Claude Code for Pro, Max, and Team plans, replacing per-command approval with autonomous runs that flag only dangerous commands. A CVSS 9.4 vulnerability (CVE-2026-54316) presented at Black Hat USA showed a single crafted GitHub issue could exfiltrate CI credentials through Claude Code's GitHub Action. Self-hosted runners also entered public beta, letting Team and Enterprise orgs run Claude Code sessions entirely on their own compute.

Headline

Auto mode is now the default in Claude Code

Anthropic announced Aug 7 that Claude Code will run auto mode by default for Pro, Max, and Team plans, with the rollout starting the following week. Instead of approving every command, you supervise longer autonomous runs while a safety layer intercepts dangerous commands — Anthropic's pitch is that constant approval prompts train developers to click "yes" without reading, and auto mode catches more real risk. You can switch back per-project or globally (see the permission modes docs). A companion post shows how Nuro, Gusto, and Garner Health run auto mode in production.

Skill of the Week

grilling by Matt Pocock

This week's editor pick. Instead of Claude agreeing with your plan and starting to build, grilling makes it do the opposite: relentlessly interrogate the plan first — assumptions, edge cases, failure modes, things you haven't thought about — before a single line gets written. From Matt Pocock, the TypeScript educator behind Total TypeScript and AI Hero. Cheap insurance against a day of building the wrong thing.

Community Releases

Expo Skills

Official agent skills from the Expo team that teach Claude how to drive Expo CLI, plus an MCP server for triggering EAS builds and checking project status from inside a session. Docs at docs.expo.dev/skills.

OpenPanel MCP Server

The open-source analytics platform now ships an MCP server with 38 tools for querying user and event data from Claude or Cursor — ask your analytics questions in the session where you're already working.

Obsidian Skills Pack by kepano

Obsidian's CEO maintains a skills pack for working with Obsidian vaults — note summarization, article extraction, vault automation. It made the rounds again this week as more people wire Claude into their notes.

1. Claude Code & CLI

Self-hosted runners for Claude Code enter public beta

claude self-hosted-runner is now in public beta for Team and Enterprise plans. It turns your own machines or containers into the execution layer for Claude Code web, mobile, and desktop sessions — keeping code and toolchains inside your network. Two modes: fixed runners (always-on machines) and on-demand runners (spun up per session). Not available for ZDR orgs.

Dynamic Workflows & `/effort ultracode`

The /workflows command (or a natural-language workflow description) now generates reusable orchestration scripts that fan out tasks to parallel subagents. /effort ultracode sets Claude to maximum reasoning for complex, large-scale work. Generated workflow scripts persist to disk so you can re-execute the same multi-agent pipeline without re-prompting — audit, migration, and cross-checked research are the headline use cases documented.

Skill and Plugin Security Scanning (Beta)

Enterprise plans can now enable automatic security scanning of third-party skills and plugins at upload/edit time. Flagged Aug 6 in the official release notes. If you're an Enterprise admin running a Claude Code deployment with community skills, turn this on before your security team asks you why you haven't.

2. Security

CVE-2026-54316: a GitHub issue could exfiltrate CI credentials via Claude Code's GitHub Action

Research presented at Black Hat USA 2026 showed that a crafted GitHub issue could reach CI through the Claude Code GitHub Action and exfiltrate repository credentials — rated CVSS 9.4 (NVD entry). Gemini CLI was hit by a parallel flaw in the same research. If you run Claude Code in GitHub Actions, update the action and audit which secrets your workflows expose to agent-triggered runs.

3. Competitor Dev Tools

Cursor — Google Workspace Plugins (Aug 3)

Cursor launched Gmail, Google Drive, and Google Calendar plugins that give coding agents direct read/write access to Workspace data from inside the editor. The underlying Google MCP servers are in Developer Preview. Relevant context if you're evaluating Claude Code vs. Cursor for teams with heavy Google Workspace footprints.

Codex CLI v0.147.0 — Agent Plugins + MCP 2026-07-28 (Aug 7)

OpenAI's Codex CLI ships portable Agent Plugins with cross-catalog search, support for the MCP 2026-07-28 protocol (paginated discovery, multi-round requests), Amazon Bedrock caching, and the ability to import Cursor-managed skills. The --approve-for-me flag enables automatically reviewed approvals. Builders evaluating multi-tool agentic setups should note Codex CLI is closing the plugin parity gap with Claude Code's marketplace.

4. Enterprise & Business

Inference hooks: inline data loss prevention for Claude Enterprise

Announced Aug 5. Inference hooks let compliance teams inspect and enforce policy on every prompt and tool-call response before they reach Claude — across Claude Enterprise surfaces including chat, Claude Code, and Cowork. If your org blocked Claude rollout on DLP grounds, this is the feature to re-open that conversation with.

Tino Cuéllar Joins Anthropic as Chief Global Affairs Officer

Former California Supreme Court Justice and ex-Carnegie Endowment president Mariano-Florentino (Tino) Cuéllar joins as Anthropic's first-ever Chief Global Affairs Officer, reporting to President Daniela Amodei. The hire coincided with Anthropic meeting the White House to discuss a voluntary AI safety testing framework — builders on regulated deployments should watch for policy guidance that follows from this appointment.

5. Research & Safety

Improving Fable 5's Biology Safeguards

Anthropic retrained Fable 5's biology classifier by rewriting its constitutional guidelines, incorporating expert feedback, and retraining on updated data — cutting biology-related fallbacks by roughly 85% across product surfaces (67% on Claude.ai, 55% on Cowork, 17% on Claude Code, 7% on the API). Virology, toxicology, and molecular design still route to Opus 5 fallbacks; trusted-access pathways for researchers are planned. Builders hitting biology-domain refusals in their applications should retest against the current model.

6. Release Notes

Claude Code:

v2.1.226 (Aug 8) — Bug fixes and reliability improvements
v2.1.225 (Aug 8) — Gateway spend-limit support in usage warnings; workspace trust prompt for claude agents; fixes for OAuth 401 on token refresh, MCP OAuth macOS keychain timeouts, consecutive-block limit miscounting safety-filter refusals, Remote Control session resume after large compacted conversations, and headless cross-session message delivery
v2.1.224 (Aug 7) — claude self-hosted-runner in beta; archive plugin source (zip over HTTPS with SHA-256 pinning); cross-session SendMessage and ListAgents (macOS and Linux); crossSessionInbound/dialogExpiry settings; sandbox credential masking with JWT-aware decode and AWS SigV4 re-signing; fixes for long project path session directory collisions
v2.1.223 (Aug 6) — Owner wildcard entries ("owner/*") for strictKnownMarketplaces/blockedMarketplaces; /teleport hint in cloud sessions; security fixes: Bash permission bypass via crafted commands, tab/Unicode-padded command hiding in approval dialogs, workflow sandbox dynamic import() escape, bypassPermissions agent definition ignoring org policy; fixes for gateway model discovery hiding prefixed Claude model IDs
v2.1.222 (Aug 4) — Security fix: worktree-isolated sessions could run destructive git commands against the main checkout; security fix: PreToolUse auto-allow hooks bypassed tool restrictions in background agent tasks; fixes for startup connectivity check hanging behind HTTPS proxies, /usage overattributing MCP server consumption, stream idle timeout on custom gateway base URLs

Agent SDKs (TypeScript):

v0.3.226 (Aug 8) — Parity with Claude Code v2.1.226
v0.3.225 (Aug 8) — Fixed background subagents in headless/SDK sessions never resuming after a background shell command or Monitor completed
v0.3.224 (Aug 7) — Added crossSessionInbound/dialogExpiry settings; subkind: 'peer-send-message' on task-notification; source: 'archive' plugin config; sandbox credential-masking fields (decode: 'jwt', awsPairs/sigv4); fixed long project path session directory collisions
v0.3.223 (Aug 6) — resumeDropsTurn option for safe truncating resumes; 529 overload errors now surfaced as api_error_status: 529 in result messages; headless sessions emit system/permission_denied stream events on auto-denied tool calls; usage vs modelUsage clarified in docs (use modelUsage for cost accounting)
v0.3.222 (Aug 4) — Fixed query({ sessionStore, resume }) not carrying user settings.json into resumed subprocess
v0.3.221 (Aug 4) — Improved skills option validation: malformed names now raise errors; fixed external MCP servers not connecting before the first turn in SDK sessions

Agent SDKs (Python):

v0.2.134 (Aug 8) — Bundles Claude CLI v2.1.226; pip install claude-agent-sdk==0.2.134
v0.2.133 (Aug 8) — Bundles CLI v2.1.225
v0.2.132 (Aug 7) — Bundles CLI v2.1.224
v0.2.131 (Aug 6) — Bundles CLI v2.1.223
v0.2.130 (Aug 4) — Bundles CLI v2.1.222
v0.2.129 (Aug 4) — Breaking: skill names with delimiters, control characters, wildcards, or leading / now raise ValueError at connect time; skills=["plugin:*"] → use skills="all"; security fix: crafted skill names could inject extra --allowedTools permission rules; bundles CLI v2.1.221

Action Items

Immediate:

Decide your auto mode stance before the rollout — auto mode becomes the default for Pro, Max, and Team plans; if you want per-command approval, set your permission mode explicitly now (per-project or globally).
Update Claude Code's GitHub Action and audit CI secrets — CVE-2026-54316 (CVSS 9.4) showed a crafted GitHub issue could exfiltrate CI credentials through agent-triggered workflow runs.
Update Python SDK to v0.2.129+ — breaking change: skills=["plugin:*"] and skills=["*"] now raise ValueError; replace with skills="all" before your next deploy.
Apply the v2.1.222 and v2.1.223 security fixes — worktree-session destructive git commands, PreToolUse auto-allow bypass, Bash permission bypass, Unicode-padded command hiding in approval dialogs, and workflow sandbox dynamic import() escape are all fixed in the past week's releases.

For Enterprise admins:

Enable skill and plugin security scanning — available in beta as of Aug 6; turn it on before security reviews ask why third-party skills aren't being scanned on upload.
Evaluate inference hooks for DLP — inline policy enforcement on every prompt and tool call across Claude Enterprise surfaces; unblocks compliance-gated rollouts.

For Team/Enterprise platform teams:

Plan self-hosted runner rollout — claude self-hosted-runner is in public beta; if your organization has network isolation requirements, this week's launch is the entry point; note ZDR orgs are excluded from the beta.

Want to catch up? Read every past issue of This week in Claude.

Did you like this issue?

One tap is the fastest way to shape next week.

👍 Liked it · 👎 Didn't like it

And I read every reply. Hit reply and tell me what you want more of, or less of, and it goes straight into the next issue.

Mert Duzgun